Please enable JavaScript.
Coggle requires JavaScript to display documents.
upgrade FCMNet, Based on FCMNet, in addition to counterfactuals, is there…
-
Based on FCMNet
Choose message(Perhaps selecting messages in difficult tasks can improve performance, but selecting messages in simple tasks will not have a big impact )
-
For a agent, the more similar action probability, the more difficult to make a decision, so communication is needed
add a discrete action, only the agents who choose the same action can communicate (namely, agents have the same thought) or add a continous action, the higher the similarity of the action, the greater the proportion of communication---maven(step 4)
learnable weight, using softmax probablity or top K to select communication agent-
message buffer(sampling message base on some tricks of experience replay)-用message作为buffer里其他经验（o，a,）的标记，类似fingerprint：用现在的策略（or message）来更改旧的策略从而避免multi-agent task中的不稳定
-
Produce messge
Give specific meaning to message (What do we want the message to represent)---But we don’t know what the best meaning of the message is. Maybe it’s better to learn the message automatically than to assign the specific meaning manually?
predict other agents's action probability(that is, how does the agent see other agents)
-
-
-
improve lstm（step 2）
-
-
Other recent models，I need time to read papers(aim to find a structure which the has multiple input and the different inputs have differnet meanings)
-
-
-
Combine messge(step 1)
-
If the simple combination of message fails, can we think it is because the superposition of multiple clear pictures will result in a chaotic picture, so if the dimensionality of the message is increased (further blurred picture) and then superimpose them will produce a more meaningful picture?
-
-
- in addition to counterfactuals, is there any other way to measure the amount of information in messages? such as entropy?
-
-
-
1.Maybe we can use something from information theory to increase mathematical support？-ok , Not only can be used to design the structure, but also can be used to analyze the meaning of the message
3.What other communication skills in real life can be achieved through neural networks--1.1consensus(localy joint action)(step 5)-now just condition on information of other agents not the policy(action) of other agents 1.2 Pheromence-- leaving trail in envroments(Formic)